Performance Trade-Offs in Using NVRAM Write Buffer for Flash Memory-Based Storage Devices

Transcription

1 Performance Trade-Offs in Using NVRAM Write Buffer for Flash Memory-Based Storage Devices Sooyong Kang, Sungmin Park, Hoyoung Jung, Hyoki Shim, and Jaehyuk Cha IEEE TRANSACTIONS ON COMPUTERS, VOL. 8, NO., JUNE 009

2 Outline Introduction Backgrounds NVRAM write buffer management polices Performance evaluation Optimistic FTL Conclusion

3 Introduction(/) Flash memory has a few drawbacks, such as the asymmetric speed of read and write operations, inability to in-place updates, very slow erasure operation, among others. One of the approaches to achieve the objective is by exploiting the buffer cache in volatile memory to delay write operations.

4 Introduction(/) Using nonvolatile random access memory (NVRAM) as a write buffer for a slow storage device has long been an active research area. For the past decade, next-generation nonvolatile memory has been under active development.

5 Introduction(/) In this paper, we suggest the utilization of smallsized, next-generation NVRAM as a write buffer to improve the overall performance. DRAM We propose a novel write buffer-aware flash translation layer algorithm, optimistic FTL, which is designed to harmonize well with SSD NVRAM write buffers. Write buffer

6 Backgrounds(/) Characteristics of the NAND Flash Memory Three basic operations Few drawbacks :. Asymmetric operations speed. Inability to in-place update. Limited lifetime. Random page write prohibition within a block

7 Backgrounds(/) Flash Translation Layer (FTL) It provides a few core functionalities such as address mapping, bad block management, and ECC check. The overall performance of the flash memorybased storage system highly depends on the mapping scheme. 7

8 Backgrounds(/) Log Block-Based Address Mapping There are two representative log block schemes: BAST and FAST. When there is no available log block, they select a victim log block and merge it with its corresponding data block(s) a merge operation. 8

9 Backgrounds(/) Three different forms of merge operations : Switch merge Partial merge Full merge data 0 log Full merge : data 0 erase + page copy data 0 Switch merge : log data 0 data 0 log 0 data 0 erase + 0 page copy Partial merge : erase + n page copy n 9

10 NVRAM write buffer management polices for flash memory Least Recently Used Page (LRU-P) Policy (page-level management) Least Recently Used Cluster (LRU-C) Policy Largest Cluster (LC) Policy Cold and Largest Cluster (CLC) Policy 0

11 NVRAM write buffer management polices for flash memory data A C log L A C Full merge FTL(BAST) A A A A B L B B B B B

12 data A A C log data L C A A switch merge L FTL(BAST) A A A B A L B B B B B

13 NVRAM write buffer management polices for flash memory Least Recently Used Page (LRU-P) Policy The replacement unit is 9 a 8 page and the least recently used (written) page in the buffer is selected as a 8 9 spaces victim. MRU Least Recently Used Cluster (LRU-C) Policy The replacement unit is 9 a 8 page cluster and the least recently accessed cluster is selected as a victim. 8 9 LRU spaces MRU LRU

14 NVRAM write buffer management polices for flash memory Largest Cluster (LC) Policy The replacement unit is a page cluster and the page cluster with the largest cluster size is selected as a victim. Largest

15 NVRAM write buffer management polices for flash memory Cold and Largest Cluster (CLC) Policy

16 Evaluation Storage device : Samsung K9NBG08UA -Gbit large block flash memory Using BAST and FAST I/O trace : FAT and NTFS

17 a) Page hit ratio b) number of destaged clusters c) average size of victim clusters 7

18 Performance Comparison : FAT trace BAST: number of log blocks = FAST: number of log blocks = 8

19 Performance Comparison : NTFS trace BAST: number of log blocks = FAST: number of log blocks = 9

20 Performance of CLC policy CLC = α *LRUC +(-α)lc α =0. 0

21 Optimistic FTL Log block management : a) Append operation b) Data block switch operation c) Log block switch operation N B : the block size in number of pages LPI stores the index of the last page stored in the log block. I min and I max be the smallest and largest page index in the victim cluster, respectively.

22 Append operation When LPI < I min : No erase operation is needed. Data block log block Data block log block 7 8 Victim cluster 7 8 LPI = ; I min = ; I max = ; LPI = ;

23 Log block switch operation(/) When I min LPI : Update map Data block log block Free block Data block log block 7 8 Erase Victim cluster 7 8 Free block LPI = ; I min = ; I max = ; LPI = ;

24 Log block switch operation(/) When I max LPI : Update map Data block log block Free block Data block log block 7 8 Erase Free block LPI = ; I min = ; I max = ; Victim cluster Extra cost : 7 8 LPI (I max - I min +) = LPI - I max +I min - LPI = ;

25 Data block switch operation When I max LPI : Update map Data block 7 8 Erase Free block log block Free block LPI = ; I min = ; I max = ; Victim cluster Update map Data block 7 8 log block Extra cost : N B - LPI +I min - LPI = ;

26 Merge latencies in each FTL algorithm BAST FAST Optimistic FTL

27 Extra overhead in each FTL algorithm. FAT trace NTFS trace 7

28 Conclusion The CLC policy not only exploits the temporal locality but also maximizes the number of simultaneously destaged pages. Simulation results have shown that the CLC policy outperforms traditional pagelevel LRU policy (LRU-P) by a maximum of percent. 8